Violence Classification Using Support Vector Machine and Deep Transfer Learning Feature Extraction

Abstract

Violence detection research is still quite a challenge for researchers and a considerable amount of effort. Before the video can be processed for classification, feature extraction is an important process to obtain important information. Determination of feature extraction and classification algorithms is an important factor for accurate classification results. This study uses deep transfer learning for feature extraction and combining it with the Support Vector Machine (SVM) classifier. The deep transfer learning algorithm in this study is a pre-trained model of Visual Geometry Group Network-16 (VGGNet-16). The video data was extracted using VGGNet-16 and then classified using SVM. Tests were carried out with 5-fold cross-validation with a variety of linear kernel, RBF, and Polynomial functions. The results were also compared with the Principal Component Analysis (PCA) feature extraction algorithm combining with SVM also. The results showed that the combination of deep transfer learning with SVM linear kernel functions resulted in higher accuracy compared to RBF and Polynomial kernel functions, and also compared to PCA combined with SVM.